CN-121981316-A - Traffic event prediction method, prediction device and prediction system

CN121981316ACN 121981316 ACN121981316 ACN 121981316ACN-121981316-A

Abstract

The invention relates to the technical field of traffic control, in particular to a traffic event prediction method, a prediction device and a prediction system, which solve the technical problem of lower prediction precision when predicting traffic events by single road information in the prior art. According to the traffic event prediction method provided by the invention, the model parameter weights of the traffic event prediction models after training are obtained by aggregation according to the model parameter weights of the traffic event prediction models of a plurality of clients, the prediction models of the traffic event prediction by the clients are integrated, the length of the prediction models corresponding to the clients can be absorbed, the short plates of the prediction models corresponding to the clients are supplemented, data and technical support are provided for the traffic event prediction models of the clients, the prediction accuracy of traffic events is improved, and secondary accidents and the expansion of a single accident range are avoided. In addition, only the model parameter weights are safely aggregated, so that the original data privacy of each client is protected.

Inventors

Yang Fengman
YI QIAN
SONG XIANGHUI
WANG DONGZHU
LI YAMENG
LIU NAN
SUN LING
Gao Zhuomiao
Jia Youfang
Ji meichen

Assignees

交通运输部公路科学研究所

Dates

Publication Date: 20260505
Application Date: 20251219

Claims (10)

1. A traffic event prediction method, comprising: Performing iterative training on a local traffic event prediction model according to first model parameter weights sent by a plurality of clients and second model parameter weights of the local traffic event prediction model until the training of the local traffic event prediction model is completed, wherein the first model parameter weights are model parameter weights obtained after the clients train the traffic event prediction model at the clients according to the received second model parameter weights; And inputting the traffic data into the traffic event prediction model after training for calculation, and generating traffic event prediction data.
2. The prediction method according to claim 1, wherein the iterative training of the local traffic event prediction model according to the first model parameter weights sent by the plurality of clients and the second model parameter weights of the local traffic event prediction model until the training of the local traffic event prediction model is completed comprises: According to the second model parameter weights of the local traffic event prediction model, the first model parameter weights sent by the plurality of clients are aggregated to update the second model parameter weights of the local traffic event prediction model; Training a local traffic event prediction model according to the updated second model parameter weight, and calculating a loss function; when the loss function meets a preset condition, training of a local traffic event prediction model is completed; When the loss function does not meet the preset condition, updating the second model parameter weight according to the model parameter weight obtained by training the local traffic event prediction model according to the updated second model parameter weight; transmitting the updated second model parameter weights to a plurality of clients so that the clients train the traffic event prediction model at the clients according to the updated second model parameter weights to obtain first model parameter weights obtained after the traffic event prediction model at the clients is trained; And returning to the step of aggregating the first model parameter weights sent by the plurality of clients according to the second model parameter weights of the local traffic event prediction model so as to update the second model parameter weights of the local traffic event prediction model.
3. The prediction method according to claim 2, wherein updating the second model parameter weights of the local traffic event prediction model according to the second model parameter weights of the local traffic event prediction model and the first model parameter weights transmitted by the plurality of clients comprises: Respectively calculating the similarity between the second model parameter weight and the first model parameter weights of the local traffic event prediction model; determining a first model parameter weight corresponding to the similarity larger than a first preset similarity as a target aggregation weight; aggregating at least one target aggregation weight to generate a new model parameter weight; and updating the second model parameter weight according to the new model parameter weight.
4. A prediction method according to claim 3, characterized in that the prediction method further comprises: Determining a client corresponding to a first model parameter weight corresponding to a similarity smaller than a second preset similarity as a marking client, and updating a marking value of the marking client; The second preset similarity is smaller than the first preset similarity.
5. The prediction method according to claim 4, characterized in that the prediction method further comprises: And when the updated marking value of the marking client is larger than a marking threshold value, storing the client information of the marking client in a refused communication list.
6. The method of predicting according to claim 2, wherein said sending the second model parameter weights to a plurality of clients comprises: Encrypting the second model parameter weight to generate a second encrypted file, wherein the second encrypted file comprises the second model parameter weight and a second key pair; and respectively sending the second encrypted files to a plurality of clients.
7. The prediction method according to claim 1, characterized in that the prediction method further comprises: Receiving a first encrypted file sent by a client, wherein the first encrypted file comprises a first model parameter weight and a first key pair, and the first model parameter weight is a model parameter weight obtained by training a traffic event prediction model at the client according to the received second model parameter weight; decrypting the first encrypted file according to the first key pair to obtain a first model parameter weight sent by the client.
8. A traffic event prediction device, comprising: A first traffic event prediction model; The first iterative training module is used for carrying out iterative training on the first traffic event prediction model according to first model parameter weights sent by a plurality of clients and second model parameter weights of the local traffic event prediction model until the first traffic event prediction model is trained, wherein the first model parameter weights are model parameter weights obtained after the clients train the traffic event prediction model at the clients according to the received second model parameter weights.
9. A traffic event prediction system, comprising: the traffic event prediction device of claim 8; The traffic event predictors are respectively installed at the plurality of clients and are in communication connection with the traffic event predicting device; wherein the traffic event predictor comprises: a second traffic event prediction model; The second iterative training module is used for receiving the second model parameter weight sent by the traffic event prediction device, training the second traffic event prediction model according to the second model parameter weight, and sending the first model parameter weight obtained after training the second traffic event prediction model to the traffic event prediction device.
10. The traffic event prediction system according to claim 9, characterized in that, The traffic event predictor also comprises a first updating module, a second updating module and a second updating module, wherein the first updating module is used for updating the first model parameter weight according to the model parameter weight obtained after training the second traffic event prediction model; the traffic event prediction device further comprises a second updating module, wherein the second updating module is used for updating the second model parameter weight according to the model parameter weight obtained after training the first traffic event prediction model.

Description

Traffic event prediction method, prediction device and prediction system Technical Field The invention relates to the technical field of traffic control, in particular to a traffic event prediction method, a prediction device and a prediction system. Background The reasons for the road traffic accidents are complex, and the reasons relate to various factors such as meteorological environment, road traffic conditions, driving behaviors, urban construction facilities, emergency measures and the like, so that the important points of attention of various departments in the prediction of the road traffic accidents are different, for example, the important point of traffic is an accident handling mode, the important point of traffic is urban construction infrastructure, and the important point of weather is severe weather. Because the emphasis points are different, the road information acquired by each department is different, but the reasons of the road traffic accidents are complex, and the prediction accuracy is lower when the traffic accidents are predicted by single road information. Disclosure of Invention The invention aims to provide a traffic event prediction method, a prediction device and a prediction system, which solve the technical problem of lower prediction precision when predicting traffic events by single road information in the prior art. As a first aspect of the present invention, the present invention provides a traffic event prediction method, including: Performing iterative training on a local traffic event prediction model according to first model parameter weights sent by a plurality of clients and second model parameter weights of the local traffic event prediction model until the training of the local traffic event prediction model is completed, wherein the first model parameter weights are model parameter weights obtained after the clients train the traffic event prediction model at the clients according to the received second model parameter weights; And inputting the traffic data into the traffic event prediction model after training for calculation, and generating traffic event prediction data. In an embodiment of the present invention, the performing iterative training on the local traffic event prediction model according to the first model parameter weights sent by the plurality of clients and the second model parameter weights of the local traffic event prediction model until the training on the local traffic event prediction model is completed includes: According to the second model parameter weights of the local traffic event prediction model, the first model parameter weights sent by the plurality of clients are aggregated to update the second model parameter weights of the local traffic event prediction model; Training a local traffic event prediction model according to the updated second model parameter weight, and calculating a loss function; when the loss function meets a preset condition, training of a local traffic event prediction model is completed; When the loss function does not meet the preset condition, updating the second model parameter weight according to the model parameter weight obtained by training the local traffic event prediction model according to the updated second model parameter weight; transmitting the updated second model parameter weights to a plurality of clients so that the clients train the traffic event prediction model at the clients according to the updated second model parameter weights to obtain first model parameter weights obtained after the traffic event prediction model at the clients is trained; And returning to the step of aggregating the first model parameter weights sent by the plurality of clients according to the second model parameter weights of the local traffic event prediction model so as to update the second model parameter weights of the local traffic event prediction model. In an embodiment of the present invention, updating the second model parameter weight of the local traffic event prediction model according to the second model parameter weight of the local traffic event prediction model and the first model parameter weights sent by the plurality of clients includes: Respectively calculating the similarity between the second model parameter weight and the first model parameter weights of the local traffic event prediction model; determining a first model parameter weight corresponding to the similarity larger than a first preset similarity as a target aggregation weight; aggregating at least one target aggregation weight to generate a new model parameter weight; and updating the second model parameter weight according to the new model parameter weight. In an embodiment of the present invention, the prediction method further includes: Determining a client corresponding to a first model parameter weight corresponding to a similarity smaller than a second preset similarity as a marking client, and updating a marking value o